tcp: skip cwnd moderation in TCP_CA_Open in tcp_try_to_open
[linux-2.6.git] / net / ipv4 / ip_fragment.c
index 8d7f107..fdaabf2 100644 (file)
@@ -4,11 +4,9 @@
  *             interface as the means of communication with the user level.
  *
  *             The IP fragmentation functionality.
- *             
- * Version:    $Id: ip_fragment.c,v 1.59 2002/01/12 07:54:56 davem Exp $
  *
  * Authors:    Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
- *             Alan Cox <Alan.Cox@linux.org>
+ *             Alan Cox <alan@lxorguk.ukuu.org.uk>
  *
  * Fixes:
  *             Alan Cox        :       Split from ip.c , see ip_input.c for history.
 #include <linux/netdevice.h>
 #include <linux/jhash.h>
 #include <linux/random.h>
+#include <linux/slab.h>
+#include <net/route.h>
+#include <net/dst.h>
 #include <net/sock.h>
 #include <net/ip.h>
 #include <net/icmp.h>
 #include <net/checksum.h>
 #include <net/inetpeer.h>
+#include <net/inet_frag.h>
 #include <linux/tcp.h>
 #include <linux/udp.h>
 #include <linux/inet.h>
 #include <linux/netfilter_ipv4.h>
+#include <net/inet_ecn.h>
 
 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
  * as well. Or notify me, at least. --ANK
  */
 
-/* Fragment cache limits. We will commit 256K at one time. Should we
- * cross that limit we will prune down to 192K. This should cope with
- * even the most extreme cases without allowing an attacker to measurably
- * harm machine performance.
- */
-int sysctl_ipfrag_high_thresh = 256*1024;
-int sysctl_ipfrag_low_thresh = 192*1024;
-
-int sysctl_ipfrag_max_dist = 64;
-
-/* Important NOTE! Fragment queue must be destroyed before MSL expires.
- * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
- */
-int sysctl_ipfrag_time = IP_FRAG_TIME;
+static int sysctl_ipfrag_max_dist __read_mostly = 64;
 
 struct ipfrag_skb_cb
 {
@@ -70,156 +60,142 @@ struct ipfrag_skb_cb
        int                     offset;
 };
 
-#define FRAG_CB(skb)   ((struct ipfrag_skb_cb*)((skb)->cb))
+#define FRAG_CB(skb)   ((struct ipfrag_skb_cb *)((skb)->cb))
 
 /* Describe an entry in the "incomplete datagrams" queue. */
 struct ipq {
-       struct hlist_node list;
-       struct list_head lru_list;      /* lru list member                      */
+       struct inet_frag_queue q;
+
        u32             user;
-       u32             saddr;
-       u32             daddr;
-       u16             id;
+       __be32          saddr;
+       __be32          daddr;
+       __be16          id;
        u8              protocol;
-       u8              last_in;
-#define COMPLETE               4
-#define FIRST_IN               2
-#define LAST_IN                        1
-
-       struct sk_buff  *fragments;     /* linked list of received fragments    */
-       int             len;            /* total length of original datagram    */
-       int             meat;
-       spinlock_t      lock;
-       atomic_t        refcnt;
-       struct timer_list timer;        /* when will this queue expire?         */
-       struct timeval  stamp;
+       u8              ecn; /* RFC3168 support */
        int             iif;
        unsigned int    rid;
        struct inet_peer *peer;
 };
 
-/* Hash table. */
-
-#define IPQ_HASHSZ     64
-
-/* Per-bucket lock is easy to add now. */
-static struct hlist_head ipq_hash[IPQ_HASHSZ];
-static DEFINE_RWLOCK(ipfrag_lock);
-static u32 ipfrag_hash_rnd;
-static LIST_HEAD(ipq_lru_list);
-int ip_frag_nqueues = 0;
+/* RFC 3168 support :
+ * We want to check ECN values of all fragments, do detect invalid combinations.
+ * In ipq->ecn, we store the OR value of each ip4_frag_ecn() fragment value.
+ */
+#define        IPFRAG_ECN_NOT_ECT      0x01 /* one frag had ECN_NOT_ECT */
+#define        IPFRAG_ECN_ECT_1        0x02 /* one frag had ECN_ECT_1 */
+#define        IPFRAG_ECN_ECT_0        0x04 /* one frag had ECN_ECT_0 */
+#define        IPFRAG_ECN_CE           0x08 /* one frag had ECN_CE */
 
-static __inline__ void __ipq_unlink(struct ipq *qp)
+static inline u8 ip4_frag_ecn(u8 tos)
 {
-       hlist_del(&qp->list);
-       list_del(&qp->lru_list);
-       ip_frag_nqueues--;
+       return 1 << (tos & INET_ECN_MASK);
 }
 
-static __inline__ void ipq_unlink(struct ipq *ipq)
-{
-       write_lock(&ipfrag_lock);
-       __ipq_unlink(ipq);
-       write_unlock(&ipfrag_lock);
-}
+/* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
+ * Value : 0xff if frame should be dropped.
+ *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
+ */
+static const u8 ip4_frag_ecn_table[16] = {
+       /* at least one fragment had CE, and others ECT_0 or ECT_1 */
+       [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
+       [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
+       [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
+
+       /* invalid combinations : drop frame */
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
+       [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
+};
+
+static struct inet_frags ip4_frags;
 
-static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
+int ip_frag_nqueues(struct net *net)
 {
-       return jhash_3words((u32)id << 16 | prot, saddr, daddr,
-                           ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
+       return net->ipv4.frags.nqueues;
 }
 
-static struct timer_list ipfrag_secret_timer;
-int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
-
-static void ipfrag_secret_rebuild(unsigned long dummy)
+int ip_frag_mem(struct net *net)
 {
-       unsigned long now = jiffies;
-       int i;
+       return atomic_read(&net->ipv4.frags.mem);
+}
 
-       write_lock(&ipfrag_lock);
-       get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
-       for (i = 0; i < IPQ_HASHSZ; i++) {
-               struct ipq *q;
-               struct hlist_node *p, *n;
+static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
+                        struct net_device *dev);
 
-               hlist_for_each_entry_safe(q, p, n, &ipq_hash[i], list) {
-                       unsigned int hval = ipqhashfn(q->id, q->saddr,
-                                                     q->daddr, q->protocol);
+struct ip4_create_arg {
+       struct iphdr *iph;
+       u32 user;
+};
 
-                       if (hval != i) {
-                               hlist_del(&q->list);
+static unsigned int ipqhashfn(__be16 id, __be32 saddr, __be32 daddr, u8 prot)
+{
+       return jhash_3words((__force u32)id << 16 | prot,
+                           (__force u32)saddr, (__force u32)daddr,
+                           ip4_frags.rnd) & (INETFRAGS_HASHSZ - 1);
+}
 
-                               /* Relink to new hash chain. */
-                               hlist_add_head(&q->list, &ipq_hash[hval]);
-                       }
-               }
-       }
-       write_unlock(&ipfrag_lock);
+static unsigned int ip4_hashfn(struct inet_frag_queue *q)
+{
+       struct ipq *ipq;
 
-       mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
+       ipq = container_of(q, struct ipq, q);
+       return ipqhashfn(ipq->id, ipq->saddr, ipq->daddr, ipq->protocol);
 }
 
-atomic_t ip_frag_mem = ATOMIC_INIT(0); /* Memory used for fragments */
+static int ip4_frag_match(struct inet_frag_queue *q, void *a)
+{
+       struct ipq *qp;
+       struct ip4_create_arg *arg = a;
+
+       qp = container_of(q, struct ipq, q);
+       return  qp->id == arg->iph->id &&
+                       qp->saddr == arg->iph->saddr &&
+                       qp->daddr == arg->iph->daddr &&
+                       qp->protocol == arg->iph->protocol &&
+                       qp->user == arg->user;
+}
 
 /* Memory Tracking Functions. */
-static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
+static void frag_kfree_skb(struct netns_frags *nf, struct sk_buff *skb)
 {
-       if (work)
-               *work -= skb->truesize;
-       atomic_sub(skb->truesize, &ip_frag_mem);
+       atomic_sub(skb->truesize, &nf->mem);
        kfree_skb(skb);
 }
 
-static __inline__ void frag_free_queue(struct ipq *qp, int *work)
+static void ip4_frag_init(struct inet_frag_queue *q, void *a)
 {
-       if (work)
-               *work -= sizeof(struct ipq);
-       atomic_sub(sizeof(struct ipq), &ip_frag_mem);
-       kfree(qp);
+       struct ipq *qp = container_of(q, struct ipq, q);
+       struct ip4_create_arg *arg = a;
+
+       qp->protocol = arg->iph->protocol;
+       qp->id = arg->iph->id;
+       qp->ecn = ip4_frag_ecn(arg->iph->tos);
+       qp->saddr = arg->iph->saddr;
+       qp->daddr = arg->iph->daddr;
+       qp->user = arg->user;
+       qp->peer = sysctl_ipfrag_max_dist ?
+               inet_getpeer_v4(arg->iph->saddr, 1) : NULL;
 }
 
-static __inline__ struct ipq *frag_alloc_queue(void)
+static __inline__ void ip4_frag_free(struct inet_frag_queue *q)
 {
-       struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
-
-       if(!qp)
-               return NULL;
-       atomic_add(sizeof(struct ipq), &ip_frag_mem);
-       return qp;
-}
-
-
-/* Destruction primitives. */
-
-/* Complete destruction of ipq. */
-static void ip_frag_destroy(struct ipq *qp, int *work)
-{
-       struct sk_buff *fp;
-
-       BUG_TRAP(qp->last_in&COMPLETE);
-       BUG_TRAP(del_timer(&qp->timer) == 0);
+       struct ipq *qp;
 
+       qp = container_of(q, struct ipq, q);
        if (qp->peer)
                inet_putpeer(qp->peer);
+}
 
-       /* Release all fragment data. */
-       fp = qp->fragments;
-       while (fp) {
-               struct sk_buff *xp = fp->next;
 
-               frag_kfree_skb(fp, work);
-               fp = xp;
-       }
-
-       /* Finally, release the queue descriptor itself. */
-       frag_free_queue(qp, work);
-}
+/* Destruction primitives. */
 
-static __inline__ void ipq_put(struct ipq *ipq, int *work)
+static __inline__ void ipq_put(struct ipq *ipq)
 {
-       if (atomic_dec_and_test(&ipq->refcnt))
-               ip_frag_destroy(ipq, work);
+       inet_frag_put(&ipq->q, &ip4_frags);
 }
 
 /* Kill ipq entry. It is not destroyed immediately,
@@ -227,48 +203,19 @@ static __inline__ void ipq_put(struct ipq *ipq, int *work)
  */
 static void ipq_kill(struct ipq *ipq)
 {
-       if (del_timer(&ipq->timer))
-               atomic_dec(&ipq->refcnt);
-
-       if (!(ipq->last_in & COMPLETE)) {
-               ipq_unlink(ipq);
-               atomic_dec(&ipq->refcnt);
-               ipq->last_in |= COMPLETE;
-       }
+       inet_frag_kill(&ipq->q, &ip4_frags);
 }
 
-/* Memory limiting on fragments.  Evictor trashes the oldest 
+/* Memory limiting on fragments.  Evictor trashes the oldest
  * fragment queue until we are back under the threshold.
  */
-static void ip_evictor(void)
+static void ip_evictor(struct net *net)
 {
-       struct ipq *qp;
-       struct list_head *tmp;
-       int work;
-
-       work = atomic_read(&ip_frag_mem) - sysctl_ipfrag_low_thresh;
-       if (work <= 0)
-               return;
-
-       while (work > 0) {
-               read_lock(&ipfrag_lock);
-               if (list_empty(&ipq_lru_list)) {
-                       read_unlock(&ipfrag_lock);
-                       return;
-               }
-               tmp = ipq_lru_list.next;
-               qp = list_entry(tmp, struct ipq, lru_list);
-               atomic_inc(&qp->refcnt);
-               read_unlock(&ipfrag_lock);
-
-               spin_lock(&qp->lock);
-               if (!(qp->last_in&COMPLETE))
-                       ipq_kill(qp);
-               spin_unlock(&qp->lock);
-
-               ipq_put(qp, &work);
-               IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
-       }
+       int evicted;
+
+       evicted = inet_frag_evictor(&net->ipv4.frags, &ip4_frags);
+       if (evicted)
+               IP_ADD_STATS_BH(net, IPSTATS_MIB_REASMFAILS, evicted);
 }
 
 /*
@@ -276,142 +223,86 @@ static void ip_evictor(void)
  */
 static void ip_expire(unsigned long arg)
 {
-       struct ipq *qp = (struct ipq *) arg;
+       struct ipq *qp;
+       struct net *net;
+
+       qp = container_of((struct inet_frag_queue *) arg, struct ipq, q);
+       net = container_of(qp->q.net, struct net, ipv4.frags);
 
-       spin_lock(&qp->lock);
+       spin_lock(&qp->q.lock);
 
-       if (qp->last_in & COMPLETE)
+       if (qp->q.last_in & INET_FRAG_COMPLETE)
                goto out;
 
        ipq_kill(qp);
 
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMTIMEOUT);
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMFAILS);
+
+       if ((qp->q.last_in & INET_FRAG_FIRST_IN) && qp->q.fragments != NULL) {
+               struct sk_buff *head = qp->q.fragments;
+               const struct iphdr *iph;
+               int err;
+
+               rcu_read_lock();
+               head->dev = dev_get_by_index_rcu(net, qp->iif);
+               if (!head->dev)
+                       goto out_rcu_unlock;
+
+               /* skb dst is stale, drop it, and perform route lookup again */
+               skb_dst_drop(head);
+               iph = ip_hdr(head);
+               err = ip_route_input_noref(head, iph->daddr, iph->saddr,
+                                          iph->tos, head->dev);
+               if (err)
+                       goto out_rcu_unlock;
+
+               /*
+                * Only an end host needs to send an ICMP
+                * "Fragment Reassembly Timeout" message, per RFC792.
+                */
+               if (qp->user == IP_DEFRAG_AF_PACKET ||
+                   (qp->user == IP_DEFRAG_CONNTRACK_IN &&
+                    skb_rtable(head)->rt_type != RTN_LOCAL))
+                       goto out_rcu_unlock;
+
 
-       if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
-               struct sk_buff *head = qp->fragments;
                /* Send an ICMP "Fragment Reassembly Timeout" message. */
-               if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
-                       icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
-                       dev_put(head->dev);
-               }
+               icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
+out_rcu_unlock:
+               rcu_read_unlock();
        }
 out:
-       spin_unlock(&qp->lock);
-       ipq_put(qp, NULL);
+       spin_unlock(&qp->q.lock);
+       ipq_put(qp);
 }
 
-/* Creation primitives. */
-
-static struct ipq *ip_frag_intern(struct ipq *qp_in)
+/* Find the correct entry in the "incomplete datagrams" queue for
+ * this IP datagram, and create new one, if nothing is found.
+ */
+static inline struct ipq *ip_find(struct net *net, struct iphdr *iph, u32 user)
 {
-       struct ipq *qp;
-#ifdef CONFIG_SMP
-       struct hlist_node *n;
-#endif
+       struct inet_frag_queue *q;
+       struct ip4_create_arg arg;
        unsigned int hash;
 
-       write_lock(&ipfrag_lock);
-       hash = ipqhashfn(qp_in->id, qp_in->saddr, qp_in->daddr,
-                        qp_in->protocol);
-#ifdef CONFIG_SMP
-       /* With SMP race we have to recheck hash table, because
-        * such entry could be created on other cpu, while we
-        * promoted read lock to write lock.
-        */
-       hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
-               if(qp->id == qp_in->id          &&
-                  qp->saddr == qp_in->saddr    &&
-                  qp->daddr == qp_in->daddr    &&
-                  qp->protocol == qp_in->protocol &&
-                  qp->user == qp_in->user) {
-                       atomic_inc(&qp->refcnt);
-                       write_unlock(&ipfrag_lock);
-                       qp_in->last_in |= COMPLETE;
-                       ipq_put(qp_in, NULL);
-                       return qp;
-               }
-       }
-#endif
-       qp = qp_in;
-
-       if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
-               atomic_inc(&qp->refcnt);
-
-       atomic_inc(&qp->refcnt);
-       hlist_add_head(&qp->list, &ipq_hash[hash]);
-       INIT_LIST_HEAD(&qp->lru_list);
-       list_add_tail(&qp->lru_list, &ipq_lru_list);
-       ip_frag_nqueues++;
-       write_unlock(&ipfrag_lock);
-       return qp;
-}
+       arg.iph = iph;
+       arg.user = user;
 
-/* Add an entry to the 'ipq' queue for a newly received IP datagram. */
-static struct ipq *ip_frag_create(struct iphdr *iph, u32 user)
-{
-       struct ipq *qp;
+       read_lock(&ip4_frags.lock);
+       hash = ipqhashfn(iph->id, iph->saddr, iph->daddr, iph->protocol);
 
-       if ((qp = frag_alloc_queue()) == NULL)
+       q = inet_frag_find(&net->ipv4.frags, &ip4_frags, &arg, hash);
+       if (q == NULL)
                goto out_nomem;
 
-       qp->protocol = iph->protocol;
-       qp->last_in = 0;
-       qp->id = iph->id;
-       qp->saddr = iph->saddr;
-       qp->daddr = iph->daddr;
-       qp->user = user;
-       qp->len = 0;
-       qp->meat = 0;
-       qp->fragments = NULL;
-       qp->iif = 0;
-       qp->peer = sysctl_ipfrag_max_dist ? inet_getpeer(iph->saddr, 1) : NULL;
-
-       /* Initialize a timer for this entry. */
-       init_timer(&qp->timer);
-       qp->timer.data = (unsigned long) qp;    /* pointer to queue     */
-       qp->timer.function = ip_expire;         /* expire function      */
-       spin_lock_init(&qp->lock);
-       atomic_set(&qp->refcnt, 1);
-
-       return ip_frag_intern(qp);
+       return container_of(q, struct ipq, q);
 
 out_nomem:
        LIMIT_NETDEBUG(KERN_ERR "ip_frag_create: no memory left !\n");
        return NULL;
 }
 
-/* Find the correct entry in the "incomplete datagrams" queue for
- * this IP datagram, and create new one, if nothing is found.
- */
-static inline struct ipq *ip_find(struct iphdr *iph, u32 user)
-{
-       __be16 id = iph->id;
-       __u32 saddr = iph->saddr;
-       __u32 daddr = iph->daddr;
-       __u8 protocol = iph->protocol;
-       unsigned int hash;
-       struct ipq *qp;
-       struct hlist_node *n;
-
-       read_lock(&ipfrag_lock);
-       hash = ipqhashfn(id, saddr, daddr, protocol);
-       hlist_for_each_entry(qp, n, &ipq_hash[hash], list) {
-               if(qp->id == id         &&
-                  qp->saddr == saddr   &&
-                  qp->daddr == daddr   &&
-                  qp->protocol == protocol &&
-                  qp->user == user) {
-                       atomic_inc(&qp->refcnt);
-                       read_unlock(&ipfrag_lock);
-                       return qp;
-               }
-       }
-       read_unlock(&ipfrag_lock);
-
-       return ip_frag_create(iph, user);
-}
-
 /* Is the fragment too far ahead to be part of ipq? */
 static inline int ip_frag_too_far(struct ipq *qp)
 {
@@ -428,10 +319,13 @@ static inline int ip_frag_too_far(struct ipq *qp)
        end = atomic_inc_return(&peer->rid);
        qp->rid = end;
 
-       rc = qp->fragments && (end - start) > max;
+       rc = qp->q.fragments && (end - start) > max;
 
        if (rc) {
-               IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
+               struct net *net;
+
+               net = container_of(qp->q.net, struct net, ipv4.frags);
+               IP_INC_STATS_BH(net, IPSTATS_MIB_REASMFAILS);
        }
 
        return rc;
@@ -441,94 +335,111 @@ static int ip_frag_reinit(struct ipq *qp)
 {
        struct sk_buff *fp;
 
-       if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time)) {
-               atomic_inc(&qp->refcnt);
+       if (!mod_timer(&qp->q.timer, jiffies + qp->q.net->timeout)) {
+               atomic_inc(&qp->q.refcnt);
                return -ETIMEDOUT;
        }
 
-       fp = qp->fragments;
+       fp = qp->q.fragments;
        do {
                struct sk_buff *xp = fp->next;
-               frag_kfree_skb(fp, NULL);
+               frag_kfree_skb(qp->q.net, fp);
                fp = xp;
        } while (fp);
 
-       qp->last_in = 0;
-       qp->len = 0;
-       qp->meat = 0;
-       qp->fragments = NULL;
+       qp->q.last_in = 0;
+       qp->q.len = 0;
+       qp->q.meat = 0;
+       qp->q.fragments = NULL;
+       qp->q.fragments_tail = NULL;
        qp->iif = 0;
+       qp->ecn = 0;
 
        return 0;
 }
 
 /* Add new segment to existing queue. */
-static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
+static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 {
        struct sk_buff *prev, *next;
+       struct net_device *dev;
        int flags, offset;
        int ihl, end;
+       int err = -ENOENT;
+       u8 ecn;
 
-       if (qp->last_in & COMPLETE)
+       if (qp->q.last_in & INET_FRAG_COMPLETE)
                goto err;
 
        if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
-           unlikely(ip_frag_too_far(qp)) && unlikely(ip_frag_reinit(qp))) {
+           unlikely(ip_frag_too_far(qp)) &&
+           unlikely(err = ip_frag_reinit(qp))) {
                ipq_kill(qp);
                goto err;
        }
 
-       offset = ntohs(skb->nh.iph->frag_off);
+       ecn = ip4_frag_ecn(ip_hdr(skb)->tos);
+       offset = ntohs(ip_hdr(skb)->frag_off);
        flags = offset & ~IP_OFFSET;
        offset &= IP_OFFSET;
        offset <<= 3;           /* offset is in 8-byte chunks */
-       ihl = skb->nh.iph->ihl * 4;
+       ihl = ip_hdrlen(skb);
 
        /* Determine the position of this fragment. */
-       end = offset + skb->len - ihl;
+       end = offset + skb->len - ihl;
+       err = -EINVAL;
 
        /* Is this the final fragment? */
        if ((flags & IP_MF) == 0) {
                /* If we already have some bits beyond end
                 * or have different end, the segment is corrrupted.
                 */
-               if (end < qp->len ||
-                   ((qp->last_in & LAST_IN) && end != qp->len))
+               if (end < qp->q.len ||
+                   ((qp->q.last_in & INET_FRAG_LAST_IN) && end != qp->q.len))
                        goto err;
-               qp->last_in |= LAST_IN;
-               qp->len = end;
+               qp->q.last_in |= INET_FRAG_LAST_IN;
+               qp->q.len = end;
        } else {
                if (end&7) {
                        end &= ~7;
                        if (skb->ip_summed != CHECKSUM_UNNECESSARY)
                                skb->ip_summed = CHECKSUM_NONE;
                }
-               if (end > qp->len) {
+               if (end > qp->q.len) {
                        /* Some bits beyond end -> corruption. */
-                       if (qp->last_in & LAST_IN)
+                       if (qp->q.last_in & INET_FRAG_LAST_IN)
                                goto err;
-                       qp->len = end;
+                       qp->q.len = end;
                }
        }
        if (end == offset)
                goto err;
 
+       err = -ENOMEM;
        if (pskb_pull(skb, ihl) == NULL)
                goto err;
-       if (pskb_trim_rcsum(skb, end-offset))
+
+       err = pskb_trim_rcsum(skb, end - offset);
+       if (err)
                goto err;
 
        /* Find out which fragments are in front and at the back of us
         * in the chain of fragments so far.  We must know where to put
         * this fragment, right?
         */
+       prev = qp->q.fragments_tail;
+       if (!prev || FRAG_CB(prev)->offset < offset) {
+               next = NULL;
+               goto found;
+       }
        prev = NULL;
-       for(next = qp->fragments; next != NULL; next = next->next) {
+       for (next = qp->q.fragments; next != NULL; next = next->next) {
                if (FRAG_CB(next)->offset >= offset)
                        break;  /* bingo! */
                prev = next;
        }
 
+found:
        /* We found where to put this one.  Check for overlap with
         * preceding fragment, and, if needed, align things so that
         * any overlaps are eliminated.
@@ -538,8 +449,10 @@ static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 
                if (i > 0) {
                        offset += i;
+                       err = -EINVAL;
                        if (end <= offset)
                                goto err;
+                       err = -ENOMEM;
                        if (!pskb_pull(skb, i))
                                goto err;
                        if (skb->ip_summed != CHECKSUM_UNNECESSARY)
@@ -547,6 +460,8 @@ static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
                }
        }
 
+       err = -ENOMEM;
+
        while (next && FRAG_CB(next)->offset < end) {
                int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
 
@@ -557,14 +472,14 @@ static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
                        if (!pskb_pull(next, i))
                                goto err;
                        FRAG_CB(next)->offset += i;
-                       qp->meat -= i;
+                       qp->q.meat -= i;
                        if (next->ip_summed != CHECKSUM_UNNECESSARY)
                                next->ip_summed = CHECKSUM_NONE;
                        break;
                } else {
                        struct sk_buff *free_it = next;
 
-                       /* Old fragmnet is completely overridden with
+                       /* Old fragment is completely overridden with
                         * new one drop it.
                         */
                        next = next->next;
@@ -572,10 +487,10 @@ static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
                        if (prev)
                                prev->next = next;
                        else
-                               qp->fragments = next;
+                               qp->q.fragments = next;
 
-                       qp->meat -= free_it->len;
-                       frag_kfree_skb(free_it, NULL);
+                       qp->q.meat -= free_it->len;
+                       frag_kfree_skb(qp->q.net, free_it);
                }
        }
 
@@ -583,50 +498,88 @@ static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 
        /* Insert this fragment in the chain of fragments. */
        skb->next = next;
+       if (!next)
+               qp->q.fragments_tail = skb;
        if (prev)
                prev->next = skb;
        else
-               qp->fragments = skb;
-
-       if (skb->dev)
-               qp->iif = skb->dev->ifindex;
-       skb->dev = NULL;
-       skb_get_timestamp(skb, &qp->stamp);
-       qp->meat += skb->len;
-       atomic_add(skb->truesize, &ip_frag_mem);
+               qp->q.fragments = skb;
+
+       dev = skb->dev;
+       if (dev) {
+               qp->iif = dev->ifindex;
+               skb->dev = NULL;
+       }
+       qp->q.stamp = skb->tstamp;
+       qp->q.meat += skb->len;
+       qp->ecn |= ecn;
+       atomic_add(skb->truesize, &qp->q.net->mem);
        if (offset == 0)
-               qp->last_in |= FIRST_IN;
+               qp->q.last_in |= INET_FRAG_FIRST_IN;
 
-       write_lock(&ipfrag_lock);
-       list_move_tail(&qp->lru_list, &ipq_lru_list);
-       write_unlock(&ipfrag_lock);
+       if (qp->q.last_in == (INET_FRAG_FIRST_IN | INET_FRAG_LAST_IN) &&
+           qp->q.meat == qp->q.len)
+               return ip_frag_reasm(qp, prev, dev);
 
-       return;
+       write_lock(&ip4_frags.lock);
+       list_move_tail(&qp->q.lru_list, &qp->q.net->lru_list);
+       write_unlock(&ip4_frags.lock);
+       return -EINPROGRESS;
 
 err:
        kfree_skb(skb);
+       return err;
 }
 
 
 /* Build a new IP datagram from all its fragments. */
 
-static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
+static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
+                        struct net_device *dev)
 {
+       struct net *net = container_of(qp->q.net, struct net, ipv4.frags);
        struct iphdr *iph;
-       struct sk_buff *fp, *head = qp->fragments;
+       struct sk_buff *fp, *head = qp->q.fragments;
        int len;
        int ihlen;
+       int err;
+       u8 ecn;
 
        ipq_kill(qp);
 
-       BUG_TRAP(head != NULL);
-       BUG_TRAP(FRAG_CB(head)->offset == 0);
+       ecn = ip4_frag_ecn_table[qp->ecn];
+       if (unlikely(ecn == 0xff)) {
+               err = -EINVAL;
+               goto out_fail;
+       }
+       /* Make the one we just received the head. */
+       if (prev) {
+               head = prev->next;
+               fp = skb_clone(head, GFP_ATOMIC);
+               if (!fp)
+                       goto out_nomem;
+
+               fp->next = head->next;
+               if (!fp->next)
+                       qp->q.fragments_tail = fp;
+               prev->next = fp;
+
+               skb_morph(head, qp->q.fragments);
+               head->next = qp->q.fragments->next;
+
+               kfree_skb(qp->q.fragments);
+               qp->q.fragments = head;
+       }
+
+       WARN_ON(head == NULL);
+       WARN_ON(FRAG_CB(head)->offset != 0);
 
        /* Allocate a new buffer for the datagram. */
-       ihlen = head->nh.iph->ihl*4;
-       len = ihlen + qp->len;
+       ihlen = ip_hdrlen(head);
+       len = ihlen + qp->q.len;
 
-       if(len > 65535)
+       err = -E2BIG;
+       if (len > 65535)
                goto out_oversize;
 
        /* Head of list must not be cloned. */
@@ -636,7 +589,7 @@ static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
        /* If the first fragment is fragmented itself, we split
         * it to two chunks: the first with data and paged part
         * and the second, holding only fragments. */
-       if (skb_shinfo(head)->frag_list) {
+       if (skb_has_frag_list(head)) {
                struct sk_buff *clone;
                int i, plen = 0;
 
@@ -645,20 +598,19 @@ static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
                clone->next = head->next;
                head->next = clone;
                skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
-               skb_shinfo(head)->frag_list = NULL;
-               for (i=0; i<skb_shinfo(head)->nr_frags; i++)
-                       plen += skb_shinfo(head)->frags[i].size;
+               skb_frag_list_init(head);
+               for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
+                       plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
                clone->len = clone->data_len = head->data_len - plen;
                head->data_len -= clone->len;
                head->len -= clone->len;
                clone->csum = 0;
                clone->ip_summed = head->ip_summed;
-               atomic_add(clone->truesize, &ip_frag_mem);
+               atomic_add(clone->truesize, &qp->q.net->mem);
        }
 
        skb_shinfo(head)->frag_list = head->next;
-       skb_push(head, head->data - head->nh.raw);
-       atomic_sub(head->truesize, &ip_frag_mem);
+       skb_push(head, head->data - skb_network_header(head));
 
        for (fp=head->next; fp; fp = fp->next) {
                head->data_len += fp->len;
@@ -668,80 +620,253 @@ static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
                else if (head->ip_summed == CHECKSUM_COMPLETE)
                        head->csum = csum_add(head->csum, fp->csum);
                head->truesize += fp->truesize;
-               atomic_sub(fp->truesize, &ip_frag_mem);
        }
+       atomic_sub(head->truesize, &qp->q.net->mem);
 
        head->next = NULL;
        head->dev = dev;
-       skb_set_timestamp(head, &qp->stamp);
+       head->tstamp = qp->q.stamp;
 
-       iph = head->nh.iph;
+       iph = ip_hdr(head);
        iph->frag_off = 0;
        iph->tot_len = htons(len);
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
-       qp->fragments = NULL;
-       return head;
+       iph->tos |= ecn;
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMOKS);
+       qp->q.fragments = NULL;
+       qp->q.fragments_tail = NULL;
+       return 0;
 
 out_nomem:
-       LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
+       LIMIT_NETDEBUG(KERN_ERR "IP: queue_glue: no memory for gluing "
                              "queue %p\n", qp);
+       err = -ENOMEM;
        goto out_fail;
 out_oversize:
        if (net_ratelimit())
-               printk(KERN_INFO
-                       "Oversized IP packet from %d.%d.%d.%d.\n",
-                       NIPQUAD(qp->saddr));
+               printk(KERN_INFO "Oversized IP packet from %pI4.\n",
+                       &qp->saddr);
 out_fail:
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
-       return NULL;
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMFAILS);
+       return err;
 }
 
 /* Process an incoming IP datagram fragment. */
-struct sk_buff *ip_defrag(struct sk_buff *skb, u32 user)
+int ip_defrag(struct sk_buff *skb, u32 user)
 {
-       struct iphdr *iph = skb->nh.iph;
        struct ipq *qp;
-       struct net_device *dev;
-       
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
+       struct net *net;
 
-       /* Start by cleaning up the memory. */
-       if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
-               ip_evictor();
+       net = skb->dev ? dev_net(skb->dev) : dev_net(skb_dst(skb)->dev);
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMREQDS);
 
-       dev = skb->dev;
+       /* Start by cleaning up the memory. */
+       if (atomic_read(&net->ipv4.frags.mem) > net->ipv4.frags.high_thresh)
+               ip_evictor(net);
 
        /* Lookup (or create) queue header */
-       if ((qp = ip_find(iph, user)) != NULL) {
-               struct sk_buff *ret = NULL;
-
-               spin_lock(&qp->lock);
+       if ((qp = ip_find(net, ip_hdr(skb), user)) != NULL) {
+               int ret;
 
-               ip_frag_queue(qp, skb);
+               spin_lock(&qp->q.lock);
 
-               if (qp->last_in == (FIRST_IN|LAST_IN) &&
-                   qp->meat == qp->len)
-                       ret = ip_frag_reasm(qp, dev);
+               ret = ip_frag_queue(qp, skb);
 
-               spin_unlock(&qp->lock);
-               ipq_put(qp, NULL);
+               spin_unlock(&qp->q.lock);
+               ipq_put(qp);
                return ret;
        }
 
-       IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
+       IP_INC_STATS_BH(net, IPSTATS_MIB_REASMFAILS);
        kfree_skb(skb);
-       return NULL;
+       return -ENOMEM;
+}
+EXPORT_SYMBOL(ip_defrag);
+
+struct sk_buff *ip_check_defrag(struct sk_buff *skb, u32 user)
+{
+       const struct iphdr *iph;
+       u32 len;
+
+       if (skb->protocol != htons(ETH_P_IP))
+               return skb;
+
+       if (!pskb_may_pull(skb, sizeof(struct iphdr)))
+               return skb;
+
+       iph = ip_hdr(skb);
+       if (iph->ihl < 5 || iph->version != 4)
+               return skb;
+       if (!pskb_may_pull(skb, iph->ihl*4))
+               return skb;
+       iph = ip_hdr(skb);
+       len = ntohs(iph->tot_len);
+       if (skb->len < len || len < (iph->ihl * 4))
+               return skb;
+
+       if (ip_is_fragment(ip_hdr(skb))) {
+               skb = skb_share_check(skb, GFP_ATOMIC);
+               if (skb) {
+                       if (pskb_trim_rcsum(skb, len))
+                               return skb;
+                       memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
+                       if (ip_defrag(skb, user))
+                               return NULL;
+                       skb->rxhash = 0;
+               }
+       }
+       return skb;
+}
+EXPORT_SYMBOL(ip_check_defrag);
+
+#ifdef CONFIG_SYSCTL
+static int zero;
+
+static struct ctl_table ip4_frags_ns_ctl_table[] = {
+       {
+               .procname       = "ipfrag_high_thresh",
+               .data           = &init_net.ipv4.frags.high_thresh,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec
+       },
+       {
+               .procname       = "ipfrag_low_thresh",
+               .data           = &init_net.ipv4.frags.low_thresh,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec
+       },
+       {
+               .procname       = "ipfrag_time",
+               .data           = &init_net.ipv4.frags.timeout,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec_jiffies,
+       },
+       { }
+};
+
+static struct ctl_table ip4_frags_ctl_table[] = {
+       {
+               .procname       = "ipfrag_secret_interval",
+               .data           = &ip4_frags.secret_interval,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec_jiffies,
+       },
+       {
+               .procname       = "ipfrag_max_dist",
+               .data           = &sysctl_ipfrag_max_dist,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec_minmax,
+               .extra1         = &zero
+       },
+       { }
+};
+
+static int __net_init ip4_frags_ns_ctl_register(struct net *net)
+{
+       struct ctl_table *table;
+       struct ctl_table_header *hdr;
+
+       table = ip4_frags_ns_ctl_table;
+       if (!net_eq(net, &init_net)) {
+               table = kmemdup(table, sizeof(ip4_frags_ns_ctl_table), GFP_KERNEL);
+               if (table == NULL)
+                       goto err_alloc;
+
+               table[0].data = &net->ipv4.frags.high_thresh;
+               table[1].data = &net->ipv4.frags.low_thresh;
+               table[2].data = &net->ipv4.frags.timeout;
+       }
+
+       hdr = register_net_sysctl_table(net, net_ipv4_ctl_path, table);
+       if (hdr == NULL)
+               goto err_reg;
+
+       net->ipv4.frags_hdr = hdr;
+       return 0;
+
+err_reg:
+       if (!net_eq(net, &init_net))
+               kfree(table);
+err_alloc:
+       return -ENOMEM;
 }
 
-void ipfrag_init(void)
+static void __net_exit ip4_frags_ns_ctl_unregister(struct net *net)
 {
-       ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
-                                (jiffies ^ (jiffies >> 6)));
+       struct ctl_table *table;
 
-       init_timer(&ipfrag_secret_timer);
-       ipfrag_secret_timer.function = ipfrag_secret_rebuild;
-       ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
-       add_timer(&ipfrag_secret_timer);
+       table = net->ipv4.frags_hdr->ctl_table_arg;
+       unregister_net_sysctl_table(net->ipv4.frags_hdr);
+       kfree(table);
 }
 
-EXPORT_SYMBOL(ip_defrag);
+static void ip4_frags_ctl_register(void)
+{
+       register_net_sysctl_rotable(net_ipv4_ctl_path, ip4_frags_ctl_table);
+}
+#else
+static inline int ip4_frags_ns_ctl_register(struct net *net)
+{
+       return 0;
+}
+
+static inline void ip4_frags_ns_ctl_unregister(struct net *net)
+{
+}
+
+static inline void ip4_frags_ctl_register(void)
+{
+}
+#endif
+
+static int __net_init ipv4_frags_init_net(struct net *net)
+{
+       /*
+        * Fragment cache limits. We will commit 256K at one time. Should we
+        * cross that limit we will prune down to 192K. This should cope with
+        * even the most extreme cases without allowing an attacker to
+        * measurably harm machine performance.
+        */
+       net->ipv4.frags.high_thresh = 256 * 1024;
+       net->ipv4.frags.low_thresh = 192 * 1024;
+       /*
+        * Important NOTE! Fragment queue must be destroyed before MSL expires.
+        * RFC791 is wrong proposing to prolongate timer each fragment arrival
+        * by TTL.
+        */
+       net->ipv4.frags.timeout = IP_FRAG_TIME;
+
+       inet_frags_init_net(&net->ipv4.frags);
+
+       return ip4_frags_ns_ctl_register(net);
+}
+
+static void __net_exit ipv4_frags_exit_net(struct net *net)
+{
+       ip4_frags_ns_ctl_unregister(net);
+       inet_frags_exit_net(&net->ipv4.frags, &ip4_frags);
+}
+
+static struct pernet_operations ip4_frags_ops = {
+       .init = ipv4_frags_init_net,
+       .exit = ipv4_frags_exit_net,
+};
+
+void __init ipfrag_init(void)
+{
+       ip4_frags_ctl_register();
+       register_pernet_subsys(&ip4_frags_ops);
+       ip4_frags.hashfn = ip4_hashfn;
+       ip4_frags.constructor = ip4_frag_init;
+       ip4_frags.destructor = ip4_frag_free;
+       ip4_frags.skb_free = NULL;
+       ip4_frags.qsize = sizeof(struct ipq);
+       ip4_frags.match = ip4_frag_match;
+       ip4_frags.frag_expire = ip_expire;
+       ip4_frags.secret_interval = 10 * 60 * HZ;
+       inet_frags_init(&ip4_frags);
+}